Compute behind a fence: US AI export controls in 2026

From a single threshold to a layered regime #

The starting point is October 7, 2022. The Bureau of Industry and Security published an Interim Final Rule imposing license requirements on advanced computing integrated circuits sold to China, defined by performance and interconnect bandwidth thresholds. The rule banned NVIDIA A100 and H100 class parts, the equivalent class from AMD, and the tools needed to fabricate them at scale.

NVIDIA responded within weeks with H800 and A800, parts that respected the bandwidth ceilings while retaining the underlying compute. On October 17, 2023 BIS published the update. It redefined the thresholds in terms of total processing performance and performance density, eliminated the bandwidth carve out, and added a presumption of denial for advanced computing chips destined for end users headquartered in 22 named countries of concern. H800 and A800 became unshippable to China within 30 days. Singapore, Malaysia, and the United Arab Emirates moved into a gray zone where transshipment risk required licensing diligence.

The third pillar landed on December 2, 2024. BIS added high bandwidth memory stacks at HBM2e and above to controlled status, controlled 24 categories of semiconductor manufacturing equipment, and added 140 Chinese entities to the Entity List. The HBM rule was the more consequential of the two. SK Hynix, Samsung, and Micron supply the HBM that pairs with every frontier accelerator. The rule routed their China sales into the same license regime as the GPUs themselves.

January 13, 2025 closed the architectural loop. The Framework for Artificial Intelligence Diffusion, published 8 days before the Biden administration left office, segmented the world into three tiers and applied compute and weight export limits per tier. Tier 1 unrestricted, Tier 2 capped, Tier 3 banned. The architecture is the operating reality for 2026.

What the AI Diffusion Framework actually does #

The Framework groups roughly 200 jurisdictions into three buckets. Tier 1 is a closed list of 18 partners covering Australia, Belgium, Canada, Denmark, Finland, France, Germany, Ireland, Italy, Japan, the Netherlands, New Zealand, Norway, South Korea, Spain, Sweden, Taiwan, and the United Kingdom. Tier 1 entities can purchase advanced compute and host frontier model weights subject to end user diligence but without numeric caps.

Tier 2 is the middle, roughly 120 countries from Brazil and Mexico to Saudi Arabia, India, Singapore, Israel, Vietnam, Indonesia, Poland, and the Gulf states. Tier 2 receives a national compute allocation expressed in compute units, a synthetic measure aggregating installed accelerator FLOPs. Universal Verified End User and National Verified End User schemes allow specific cloud operators or strategic projects to draw against larger allocations after diligence.

Tier 3 is the closed loop: China, Russia, Belarus, Iran, North Korea, Venezuela, Cuba, Syria, plus the 22 arms embargoed destinations. Effectively no advanced compute, no closed weight frontier model export, no controlled inference endpoint access.

The mechanism hyperscalers and chipmakers care about is per shipment authorization rather than per country quota. A Tier 2 country's compute unit cap is enforced through individual license decisions on each accelerator order, denominated in H200 equivalent units. NVIDIA, AMD, and Intel each track allocation against the cap as a constraint on their forward order book.

What is allowed to ship to China and what is not #

NVIDIA A100, A800, H100, and H800 class parts cannot be shipped to China. H20, the part designed in late 2023 to fit under the October 2023 thresholds, remains shippable subject to license and shipped in volume through 2024 and 2025. Reuters reporting in 2025 placed H20 China revenue above 12 billion dollars over the year, consistent with NVIDIA's segment disclosures. Blackwell B100, B200, and GB200 class parts are subject to the December 2024 thresholds and the Diffusion Framework country review. A China specific Blackwell variant has been reported in trade press, though as of Q1 2026 BIS guidance has not formally cleared a successor to H20.

On memory, the December 2, 2024 rule made HBM2e, HBM3, and HBM3e shipments to China subject to license with a presumption of denial for advanced computing end uses. SK Hynix, the dominant HBM supplier with 50 percent plus share through 2024, paused direct China HBM sales in early 2025. Samsung and Micron followed.

On tools, the 24 controlled categories cover advanced lithography below the deep ultraviolet immersion node, etch and deposition for sub 14 nanometer logic and HBM stacking, and metrology tied to 3D NAND above 128 layers. ASML extreme ultraviolet was already banned to China. The December 2024 expansion captured a wider band of Applied Materials, Lam Research, KLA, and Tokyo Electron tools. Japan and the Netherlands aligned their controls in mid 2024, closing the parallel licensing path through allied suppliers.

CHIPS Act, TSMC Arizona, and the production wedge #

Export controls choke off where compute can go. The CHIPS and Science Act of August 2022 funds where advanced compute gets built. The two policies are intentionally paired. By Q1 2026 the production side has begun to deliver volume rather than promises.

TSMC Arizona Phase 1 began production of 4 nanometer logic in Q4 2024, the company's first leading edge fab outside Taiwan. Phase 2 targets 3 nanometer production by 2028. A Phase 3 commitment, included in the March 2025 expansion announcement that lifted total Arizona investment to roughly 165 billion dollars, targets 2 nanometer and A16 nodes by decade end. The Trump administration secured the expansion in part through tariff threats during Q1 2025 trade negotiations, documented in Reuters and Financial Times reporting through March 2025.

Intel 18A remains on track for Q3 2025 ramp at Ohio One and Arizona Fab 52, with PowerVia backside power and RibbonFET gate all around. Samsung Foundry Taylor Texas slipped from 2024 to 2026 ramp. Micron's Idaho and New York memory commitments anchor to mid decade rather than near term capacity. The CHIPS Act has shifted roughly 450 billion dollars of announced semiconductor capex into the United States, only a fraction of which has converted to wafer starts as of 2026.

The question for 2027 to 2030 is whether US and allied capacity covers frontier training demand. Today the answer is no. TSMC Hsinchu and Tainan still produce the bulk of leading edge wafers. Arizona Phase 1 plus Phase 2 plus Intel 18A plus Samsung Taylor will close roughly half the gap by 2028 on plausible ramp curves. The other half remains in Taiwan, the geopolitical asymmetry subsidy alone cannot solve.

DeepSeek and the policy theory under stress #

On January 27, 2025 DeepSeek released the R1 model, an open weight reasoning model that performed within striking distance of GPT-4 class systems on multiple public benchmarks. NVIDIA stock fell 17 percent on the day, the largest single day market capitalization loss for any US listed company on record. R1 combined mixture of experts architecture, FP8 training, group relative policy optimization, and aggressive engineering on a smaller installed base of accelerators, plausibly H800 and H20 class parts acquired before or under the current controls.

The policy implication is contested. One reading is that controls have not worked because a Chinese lab reached frontier reasoning capability anyway. The opposite reading is that controls have shaped the curve. R1 was trained on roughly an order of magnitude less compute than the comparable US frontier model, which is what a binding constraint should look like. R1 closed the gap on a one shot benchmark but has not produced the equivalent of the multi modal, agentic, long context system stack the leading US labs operate.

Frontier training in 2025 plausibly required 10 to the 26 floating point operations end to end, consistent with Epoch AI and SemiAnalysis estimates and the public statements of OpenAI, Anthropic, and Google DeepMind. That works out to roughly 50,000 H100 equivalents running for two to three months at high utilization. The compute envelope a Chinese lab can assemble inside the controls is bounded by pre 2023 stock, H20 class parts shipped under license, domestic Huawei Ascend 910B and 910C accelerators, and SMIC N+2 capacity. SemiAnalysis estimates put Huawei Ascend production at 200,000 to 400,000 units across 2024 to 2025.

The honest reading is that controls are slowing rather than blocking. Slowing has policy value if frontier capability matters time wise. It has less value if open weight diffusion eventually catches up. Both can be true. Neither side has a clean win from the January 2025 episode.

What hyperscalers, foundries, and governments should plan for #

For hyperscalers, compute distribution in 2026 is a regulated activity, not a logistics problem. AWS, Microsoft, Google, Oracle, Meta, and the major neoclouds need Tier 2 country compute allocation tracking inside the cloud control plane, transparent license status for enterprise customers, and prepared VEU and NVEU enrollments in priority Tier 2 markets. India, Singapore, Saudi Arabia, the United Arab Emirates, Israel, Brazil, and Mexico are where allocation tightness binds first. Compute scarcity becomes a contractual term rather than a commodity input.

For foundries, the binding question is the geographic footprint of leading edge capacity in 2028. TSMC has Arizona, Kumamoto, and Dresden on the diversification curve, but Hsinchu and Tainan still produce the majority of leading edge wafers. Customers should price the option value of multi sourcing across TSMC Arizona, Intel Foundry Services, Samsung Foundry, and Rapidus. Leading edge process technology requires a 5 to 7 year horizon from groundbreaking to qualified high volume manufacturing.

For governments outside the US, the Tier 2 designation is consequential. A country's allocation under the Diffusion Framework determines whether a national AI strategy is real industrial policy or a slogan. The Gulf states moved fastest. The G42 and US arrangement of 2024 was the template. India, Indonesia, Vietnam, Brazil, and Mexico now need analogous bilateral structures to lift their compute caps. Countries that secure VEU and NVEU pathways early will host the regional cloud zones emerging market AI workloads require.

For the United States, unfinished work is the export control architecture for inference endpoints, model weights, and post training data. The Trump administration AI National Action Plan signed in July 2025 reorganized BIS, NSC, and DOE coordination on this front. The April 2025 BIS internal memo on export control review, reported by Reuters, signaled a pragmatic posture: keep the architecture, refine calibration, prioritize enforcement against transshipment through Tier 2 jurisdictions.

The fence around frontier compute is real, layered across chips, memory, tools, weights, and country tiers, administered through license rather than slogan, and now a structural input to every AI strategy decision. Hyperscalers, foundries, and governments that build the fence into their plans will move faster than those that argue with it.

Sources #

• BIS Federal Register notice, Implementation of Additional Export Controls, October 7, 2022
• BIS Federal Register notice, Advanced Computing Update, October 17, 2023
• BIS Federal Register notice, HBM and Semiconductor Manufacturing Equipment, December 2, 2024
• BIS Framework for Artificial Intelligence Diffusion Interim Final Rule, January 13, 2025
• US Department of Commerce export control announcements
• NVIDIA Form 10-K and 10-Q SEC filings
• TSMC investor relations and Arizona expansion announcements
• SK Hynix investor relations and HBM disclosures
• Samsung Electronics quarterly results and foundry roadmap
• Micron Technology investor relations
• Reuters coverage of US chip export controls and DeepSeek market reaction
• Financial Times semiconductor and AI policy coverage
• SemiAnalysis research on Huawei Ascend, HBM, and frontier compute
• Epoch AI compute trends database
• CHIPS Act program and Department of Commerce funding awards